Building an Industry Chain Knowledge Graph: Theory, Architecture, and Key Methods

The talk introduces the concept of an industry‑chain knowledge graph as a multi‑dimensional tool for financial regulation, investment, and regional economic development, emphasizing its role in value discovery and risk identification across three typical scenarios: regulatory applications, market services, and government‑driven economic planning.

1. Theory and Knowledge – The industry chain is driven by the need to uncover value and assess risk in the securities market. Three categories of use cases are identified: regulatory platforms (e.g., IPO review), market services (e.g., loan and risk management), and regional development (e.g., industrial upgrading and investment attraction). The article outlines four content blocks: theory, architecture, key methods, and examples, with a macro‑engineering perspective.

2. Architecture and Process – The construction path follows four steps: data collection (announcements, research reports, news), framework design (industry → sector → company hierarchy), automated construction using NLP, and manual verification. The system architecture consists of an ontology layer (industry, sector, company), a computation layer (entity extraction, fusion, storage in graph databases), and a knowledge layer that serves downstream applications.

3. Key Methods – The pipeline relies on a suite of NLP algorithms: language models (Word2Vec, BERT pre‑trained on financial texts), lexical and syntactic analysis, document parsing (PDF → structured elements), industry‑level classification, upstream‑downstream relationship extraction (multi‑head selection), synonym detection (BPE‑based embeddings), entity alignment, and batch processing via a distributed message‑queue system.

4. Examples and Applications – Demonstrations include visualizing the industry chain (resource, manufacturing, consumption, knowledge layers), specific sector maps for new‑energy vehicles and lithium‑ion battery materials, and practical use cases such as IPO price‑comparison, product‑margin benchmarking, and investment screening for industrial robots and POCT devices.

5. Q&A – The speaker explains why both BERT and Word2Vec are retained (complementary precision vs. recall), how industry taxonomies are pre‑defined and later enriched, and the evaluation approach for industry classification (human‑in‑the‑loop iterative labeling) and multi‑label handling for companies spanning multiple sectors.

The presentation concludes with acknowledgments and community invitations.